Aerospace Remote Sensing for Domestic Crop Area Estimationby the United States Department of Agriculture

By Char les E. Caudill

INTRODUCTION

1. Background

1.1. Current Methodology for Estimating Crop Area

1.2. Aerospace Remote Sensing

2. Remote Sensing Research for Applications to Crop Area Estimation

2.1. Initial Research During 1972-77

2.1.1. Methodology

2.1.2. Development of Computing Capability

2.2. Iowa Project - 1978

3. AgRIST ARS

3.1. Departmental Program

3.2. SRS Role

4. Future Use of Aerospace Remote Sensing

SUMMARY

REFERENCES

."

Aerospace Remote Sensing for Domestic Crop Area Estimation

by the United States Department of Agriculture

By Charles E. CaudiH

Introduction

In the United States of America (US) the mission of collecting and disseminating current

statistics on the Nation's agriculture has been assigned to the Statistical Reporting

Service (SRS) in the U. S. Department of Agriculture (USDA). By tradition, and in some

instances by law, this mission includes responsibility for national and state level statistics.

Responsibility for local area (minor civil divisions such as counties, etc.) agricultural

statistics has in general been assigned to state governments. However in most states, SRS

~s been able to combine these two responsibilities through cooperative arrangements

with state government agencies. In most cases, with the State Departments of

Agriculture. SRS through its national office in Washington, D. C. and its 44 state offices

(a single office serves both Maryland and Delaware and one office serves all of the six

New England States) issues about 300 national and 9,000 state reports each year (I 7).

Comprehensive current statistics on all aspects of the Nation's agriculture are included in

these reports. Content ranges from weekly subjective assessments of weather and its

effects on agricultural production to precise quantitative estimates of agricultural

inventories, crop and livestock production, prices, marketings, inputs, wage rates, etc.

In addition to its primary mission of collecting and disseminating agricultural statistics,

SRS conducts a comprehensive program of methodological research to improve the

statistical methods used for its primary mission. The ultimate goal of the research is to

enhance the accuracy of the agricultural statistics produced and/or to reduce the costs

associated with a given level of accuracy. Improved sampling, data collection and

- I -

analysis procedures are constantly under study and development with this goal in mind.

New technology, such as computers and aerospace remote sensing, as well as applications

of new crop production practices constantly require study and research to understand how

they affect or can be used to improve agricultural statistics. Since some of the most

important statistics produced by the SRS are the estimates of area planted to various

crops a significant proportion of the agency's research resources have always been

devoted to improving the methods and data used for these estimates. During the past

decade this researc::h has been concentrated on the development of methodology to use

remote sensing data collected by earth-orbiting satellites for making crop area estimates.

This research will be discussed in detail. First the present operational procedures used for

making crop area estimates are briefly described. Then a description of remote sensing

technology and the research conducted over the past decade will be discussed. Finally

results of using aerospace remote sensing data for a large scale application test in 1978

are presented and prospects for the future examined • .\.

1. Background

1.1. Current Methodology for Estimating Crop Area (17)

SRS estimates of crop area are developed from data collected through sample surveys.

Generally these surveys are based on relatively small samples because of costs, and

timeliness requirements for data summarization, analysis, and publication (typically the

time interval from the start of data collection until estimates are published is about one

month). Since 1965 national and state level estimates of crop area have been based

primarily on data collected through a probability area sample, commonly referred to as

the June Enumerative Survey (JES). This survey uses a national sample of about 16,000

- 2-

areas of land. These areas of land average about one mile square and are called sampling

units or more commonly "segments." They are randomly selected from an area sampling

frame. The area sampling frame used for this purpose was developed by using maps,

published statistics, aerial mosaics or photographs, and other available materials to

stratify the entire land area of the US into major land use strata (e.g., cultivated land,

rangeland, woodland, cities and nonagricultural land). This stratification permits very

efficient samples to be designed which provide precise and independent estimates at the

state and national levels. The sample size for the JES only accounts for about one-half of

one percent of the total land area in the country, yet it produces estimates of area

planted to various crops which have sampling errors ranging from one to two percent at

the national level and three to five percent at the state level.

Data collection for the JES is done through personal interviews with individual farmers.

Interviewers who are usuaUy hired on a part-time basis are thoroughly trained before each

JES to assure uniformity of data collection methods throughout the nation. In addition to

the training program associated with the JES many other quality controls are"included in

the JES operating procedures. These include detailed instruction manuals and ~aterials,

close field supervision of interviewers, built-in questionnaire checks, and re-enumeration

of a subsample of segments by supervisors. Aerial photographs are used to locate and

identify the segments and to compare area, on a field-by-field basis, reported by farmers

with measured area on the photographs.

Interviewers use state and county road maps to locate segments and aerial photographs to

define precisely the segment and field boundaries. Each person who operates land within

the segment is interviewed to obtain information about area and use for every field within

- 3-

the segment. This information, which includes both net and gross field area for fields

used for planted crops, is carefully recorded on the survey questionnaire for later

summarization in the state office. Data collection for the JES starts during the last week

of May each year and the survey is completed in June. This includes data collection,

editing, summarization and analysis, and publication of planted area estimates.

1.2. Aerospace Remote Sensing

"Remote sensing is the acquisition of information about an object without physical

contact ..• However, the use of the term remote sensing usually refers to the gathering )

and processing of information about the earth's environment, particularly its natural and

cultural resources, through the use of photographs, and related data acquired from an

aircraft or satellite" (14, p. 1).

The term "aerospace remote sensing" is generally defined to refer to remote sensing from

space, especially from earth-orbiting dr geo-synchronous satellites. Although some\

reference to uses of conventional aerial photography will be made, in this paper the

primary purpose is to describe how aerospace remote sensing data has been and is being

used in the US to estimate land area devoted to various crops.

Specifically the paper presents how data from an earth-orbiting series of satellites

(originally called the Earth Resources Technology Satellite (ERTS), now referred to as

Landsat) is being used by SRS to estimate crop area in selected parts of the US. There

have been four Landsat satellites launched since 1972. Each of these has included a four-

channel Multispectral Scanner (MSS) sensor as part of its payload. To date, the MSS has

- 4-

provided the most useful data for SRS purposes. Therefore this paper will be limited to

describing research for developing techniques and procedures for using MSS data.

The MSS views a ground scene of approximately 185 Km by 185 Km in area, with a ground

resolution of about 80 meters. For each ground resolution element (referred to as a

picture element or "pixel") the MSS measures reflected energy in _four spectral bands

covering two regions of the electromagnetic spectrum: bands 4 and 5 in the visible range;

and bands 6 and 7 in the near-infrared. For each "pixel" the MSS data consists of four

values, one measurement in each spectral band. For each scene with 0.8 x 107 "pixels"

there are 3.2 x 107 values or measurements which must be processed for making crop area

estimates (14). This large data processing task has been the focus of much of the research

devoted to using this technology.

2. Remote Sensing Research for Applications to Crop Area Estimation

2.1. Initial Research During 1972-77

In 1972, SRS with joint sponsorship by the National Aeronautics and Space Administration

(NASA), began research to develop uses of aerospace remote sensing data collected by the

MSS on the first Landsat satellite. The objectives of this research were: to develop new

methodology for using MSS data to improve agricultural estimates; to provide knowledge

of the information content of the MSS data; and to develop computational capabilities for

extracting this information. From the very beginning of this research program emphasis

by SRS has been on a "digital" analysis approach rather than on an "image or photo-

interpretive" approach (12). Primary reasons for this include:

(a) Lack of trained staff for photo or image interpretation, with little or no prospect of

obtaining such expertise, and

-5-

(b) The requirement for timely processing and release of agricultural information.

SRS staff assigned to the remote sensing research work in 1972 consisted entirely of

statisticians and computer programmers (9). Experience and training in remote sensing

techniques would have been extremely valuable; however, it was simply not available

within SRS at that time. Since then a few staff members trained in remote sensing have

been added to the SRS staff so that a multi-disciplinary approach is possible today. Some

"image" analysis approaches particularly for land use/land cover analysis have been

developed; however,_ the predominant research is still aimed at developing more efficient

digital analysis techniques. This relates to the need for crop area estimates, whether

based on traditional ground surveys such as the JES or on remote sensing data, to be

prepared and published within a few days or few weeks after the data are collected.

2.1.1. Methodology

Even though the SRS remote sensing research was directed at producing estimates (of

area devoted to specific crops), it never had as a primary goal the complete ~ubstitution•..

of remote sensing estimates for those based on ground surveys (e.g., the JES)• .Indeed it

was readily apparent at the beginning of the research in 1972, that a more logical

approach would be to develop a methodology that would combine the strengths of the two

data sources to produce more precise estimates than either could do independently.

The data collected for the JES provided a natural "ground truth" sample, while Landsat in

effect provided a complete census of measurements (if the effect of clouds are ignored).

Consequently, research during the 1972-77 period focused on developing the necessary

techniques and procedures for integrating these two data sets (JES data with Landsat

data) to produce more precise estimates of area devoted to specific crops.

- 6-

Since the segments enumerated for the JES represent a probability sample of the entire

land area of the U.S. a major thrust of the SRS remote sensing research was to use the

"inference" capability of this data for processing and analyzing the Landsat data in order

to make statistically reliable estimates I of crop area. This included the development of

re~istration2 methods for precise matching of the common data from the two sources,

followed by use of a subsample of the matched data to estimate discriminate function

parameters for classification3 of Landsat pixels into desired land use or crop type

categories. Once this was accomplished the discriminant functions were tested by

classifying another subsample of matched pixels into categories. Results of this

classification, while never completely accurate when compared with the ground

enumerated crop area data, were sufficiently accurate to bear a close relationship with-the ground data. This relationship was used to estimate the parameters of the regression

estimator described by Cochran (5). Classification results for entire Landsat scenes or

combination of scenes were used in this estimator to make estimates of area devoted to

each crop or category of interest. Also, an effectiveness criterion for measuring the

success of using Landsat data was relatively easy to derive from this approach. This

criterion, the ratio of the variance of the crop area estimate from the ground data only,

to the variance of the crop area estimate from the Landsat regression estimator, is

generally referred to as the relative efficiency (RE). RE is the multiplier that would be

necessary to increase the sample size of the JES in order for the' estimates from the

ground data alone to be as precise as the regression estimator using both the ground and

remote sensing data. When total costs of the two approaches are considered results

currently indicate that RE's ~ 2.0 are necessary for the use of Landsat data to be cost-

effective.

-7-

2.1.2. Development of Computing Capability

While many of the early studies for developing techniques and procedures for using

Landsat data for crop area estimates were conducted by using ad-hoc computational

resources, it was apparent from the very beginning that a significant investment in data

processing, especially software, would be required. The sheer volume of data (3.2 x 107

values for each Landsat scene) plus the statistical computations involved in the

multivariate and discriminate analysis, the classification of each pixel, and the final

regression estimation of each land use or crop type category, all combined to mandate a

significant effort to develop an advanced computational capability.

Starting in 1974, SRS and the Center for Advanced Computation (CAC) at the University

of Illinois entered into a joint agreement to develop a software system to process Landsat

digital data. An interactive version of Purdue University's LARSYS (one of the first

major software systems developed in the US for processing spectral data) (13) developed

by CAC and the U. S. Department of the Interior was used as a starting point for the...proposed software system. The system (EDITOR) was redesigned to analyze SRS's ground-

gathered data from the JES, jointly with the Landsat digital data, to classify all pixels in

entire Landsat scenes, and finally to calculate the regression estimates for the crops or

land use categories of interest. The EDITOR system was designed to provide a

teleprocessing network capability, using computers of different types and sizes located at

several locations. Originally the network used included a UNIVAC 1108 on the INFONET

network, an IBM 370 at the USDA's Washington Computer Center (WCC), an IBM 370 at

CAC in Illinois, a PDP-I0 at the Bolt, Beranek, and Newman (BBN) Data Processing

Center in Cambridge, Massachusetts, and finally the ILLlAC-IV at the NASA Ames

Research Center (ARC) in California. Both the BBN computers and the ILLIAC-IV were

-8-

on the ARPANET, a teleprocessing network developed by the Department of Defense

{DOD). A detailed discussion of the initial capabilities of the EDITOR system is included

in a paper written by Ozga, M. et al in 1977 (l5). As a system for remote sensing research

EDITOR has continued to evolve into a more efficient system for operational processing.

This evolution included changes in both the EDITOR software and the computers used for

the system. The UNIVAC 1108 has been replaced by an IBM 370 system operated by

Martin Marietta Data Systems, which currently has the SRS teleprocessing contract, and

the ILLIAC-IV at ARC has been replaced by a CRAY XMP. Almost all of the algorithms

in the EDITOR software have been improved or replaced by new algorithms as these have

been developed through the SRS remote sensing research program. A partial list of these

improvements would include new algorithms for registration (2, 6), classification (3), and

statistical estimators (4, 18). Evolution of the EDITOR system is still continuing as new

hardware becomes available and the software is being converted into more easily

transportable code. Currently a major effort by SRS and ARC is underway to convert the

EDITOR code from the original ad-hoc programming languages (SAIL, RATFOR,

Assembler, etc.) used for research purposes into PASCAL. Also a new version of EDITOR

(called mini-EDITOR) is being developed for use with l6-bit microcomputers. Although

data through-put will not be as fast for a microcomputer system it may provide a cost-

effective replacement for some functions of EDITOR on the BBN computers and the

CRA Y. Positive results of remote sensing research by SRS over the past decade are

thoroughly integrated by the EDITOR system. Consequently by making the software

transportable it will be easier for SRS to share this technology with other users.

- 9-

2.2. Iowa Project - 1978

Many small and large remote sensing research studies were conducted by SRS in the 1972-

77 period to develop methodology and computing capabilities to implement the

methodology. These studies were done over agricultural areas in South Dakota, Kansas,

Missouri, Idaho, and California (10, 19). Major demonstration tests covering the entire

states of Illinois in 1975 and Kansas in 1976 (7,8) provided significant new evidence of the

usefulness of Landsat data for crop area estimation.

Based on results of previous research, SRS decided to demonstrate the EDITOR System by }

processing Landsat data for the entire state of Iowa in 1978 (11). The goal of this project

was to acquire the Landsat data during the optimum time period (late July-early

September> for separating corn (maize) and soybeans and to process the data in time for

use by the Crop Reporting Board (CRB) in establishing the final 1978 estimates for these

crops. Corn and soybeans are the two most important crops in Iowa. Together they

accounted for about 87 percent of the area planted to all crops in 1978.,"

In 1978 the entire Landsat system operated by NASA was still deemed experimental.

Because of this, real-time data delivery after satellite acquisition was not possible.

Median delivery time after acquisition of Landsat digital data was 49 days with a range

from only 32 days to as much as 93 days. Still even with this constraint the goal for the

Iowa project was met, and the CRB used the Landsat regression estimates in preparing the

Annual Crop Inventory report for 1978.

Because the Landsat satellites are in a polar orbit they pass over the US in a northeast to

southwest direction so that data are collected for a given area from east to west.

- 10 -

Figure 1 shows this for the state of Iowa in 1978. With this orbit configuration

confounded by "cloud cover" Landsat data for six different dates were required to cover

Iowa in 1978. The twelve Landsat scenes are listed in Table 1.

Table 1: Landsat Scenes Used for 1978 Iowa Project

Path Row Date Percent Cloud Scene IDCover

30 30 August 19 0 30167-1627431 August 19 0 30167-16280

29 30 August 9 0 21295-1601331 August 9 40 21295-1602032 August 18 0 30166-16224

28 30 September 4 60 30183-1616231 September 4 0 30183-1616432 September 4 0 30183-16171

27 30 August 7 10 21193-1550031 August 7 15 21293-1550232 August 7 10 21293-15505

26 31 August 6 0 21292-"15444

Even by acquiring data over as many dates as possible some areas of the state were not

included in the Landsat analysis due to cloud cover. These areas are shown in Figure 2,

which also shows the analysis districts for which regression estimates were computed in

1978. Analysis districts are formed to follow minor civil division boundaries <Counties)

which are completely covered by a Landsat scene or a Landsat pass <Consecutive scenes

on the same day). This approach is required because as the atmospheric and vegetation

conditions change from day to day reflected energy as measured by the MSS will vary for

the same areas. SRS has found that these "signatures" are not extendable over time,

- 11 -

therefore the analysis district concept has been included in the SRS remote sensing

methodology, with those districts for which no Landsat data are available being estimated

by the JES ground data only.

Crop area estimates for corn and soybeans were computed at the state, analysis district,

and individual county levels. Results are shown in Tables 2 and 3 for the state and

analysis district levels. Improvements in precision for the regression estimate (Landsat

and ground data) ~ersus the direct expansion estimate (JES ground data only) were

substantial, as measured by the relative efficiencies shown in these tables. As shown in

Figure 2, Landsat data were not available for 13 of the 99 Iowa counties during the late

July to mid-September period in 1978. For these counties estimates from the JES were

added to the other analysis district estimates to produce the total estimate for the state.

Coefficients of variation for the individual county level estimates ranged from 7 to 60

percent; however, other small area estimators for counties are being developed which will

hopefully provide county level estimates with relative mean square errors of less than 20

percent. The state level Landsat estimates were used by the USDA's Crop Reporting

Board in preparing the Annual Crop Summary for Iowa. Although not the only information

used for this summary the remote sensing input influenced the Board to change the

estimates for both corn and soybeans slightly from what the other data indicated.

Subsequent evaluation based on check data (marketings, exports, etc.) indicated that the

changes based on Landsat estimates were toward the "true" values for these crops.

- 12 -

Figure 1

LANDSAT Scene locationsPath 30 Path 29 Path 28 Path 27 Path 28

-Figure 2

Iowa Analysis Districts

-, - - --- -28

2c -3dII Data not .nelyZeci D Cloud covered

- 13 -

Table 2: Estimates of Corn Area Planted, Coefficients of Variation, andRelative Efficiencies by Analysis District, Iowa, 1978

JES Estimates Landsat Estimates

Analysis Classified RelativeDistr ict Pixels 1.1 Area Coefficient Area Coefficient Efficiency

Planted of Planted ofVariation Var iation

(1,000 (1,000 (1,000Hectares) Hectares) (Percent) Hectares) (Percent)

1 1;306 1,462 3.5 1,460 2.2 2.52A 924 829 4.5 819 2.5 3.32B 464 332 11.5 454 3.4 6.02C 125 106 11.0 110 9.5 1.23A ~/ 657 4.4 ~/ ~/ ~/3B 345 276 10.0 268 8.5 1.53C 590 551 7.5 542 6.0 1.63D 59 84 17.8 83 18.6 .94 1,059 1,030 6.7 896 4.5 3.05 132 148 11.1 150 6.0 3.3

State Total 5, 66l~/ 5,475 2.3 5 4391/ 1.5 2.4,\.

1/ Converted to hectares.2/ Landsat data not available."J/ Includes JES estimate for analysis district 3A.

- 14 -

Table 3: Estimates of Soybean Area Planted, Coefficients of Variation, andRelative Efficiencies by Analysis District, Iowa, 1978

JES Estimates Landsat EstimatesAnalysis Classified RelativeDistr ict Pixels 11 Area Coefficient Area Coefficient Efficiency

Planted of . Planted ofVariation Variation

(1 ,000 (1,000 (1,000Hectares) Hectares) (Percent) Hectares) (Percent)

1 -760 748 8.1 782 4.0 3.72A 650 655 6.8 675 3.4 3.72B 244 257 12.9 256 6.1 4.62C 94 95 25.0 97 11.7 4.43A 2:/ 402 9.2 2:/ 2:/ 2:/3B 84 87 28.0 125 9.4 4.33C 370 329 14.5 338 7.1 4.03D 79 83 32.6 96 10.2 7.64 343 441 12.7 425 8.0 2.75 35 47 29.2 49 12.5 5.1

Sta te Total 3 060~/ 3, 144 3.9 3 245~l 2.5 2.4, ,

1/>.

Converted to hectares.21 Landsat data not available.11 Includes JES estimate for analysis district 3A.

3. AgRIST ARS

Considerable detail has been provided for the 1978 Iowa project, because it produced for

the first time quasi-operational estimates based on use of Landsat data. It provided a

means of integration of much of the research completed at that time and it provided a

partial base from which the USDA could start a major new remote sensing program in

1980. This program was called "Agriculture and Resource Inventory Surveys Through

Aerospace Remote Sensing" (AgRIST ARS) (1).

- 15 -

3.1. Departmental Program

The AgRISTARS program was initiated in fiscal year 1980 in response to an initiative

issued by the USDA. Led by USDA, the program was developed as a cooperative effort

with NASA, the National Oceanic and Atmospheric Administration (NOAA) of the U.S.

Department of Commerce (USDC), the U. S. Department of the Interior (USDI), and the

U. S. Agency for International Development (AID).

The program goal \Vas to determine the usefulness, cost, and extent to which aerospace)remote sensing could be integrated into existing or future USDA systems to improve the

objectivity, reliability, timeliness, and adequacy of information required to carry out

USDA missions. The program was designed to provide research for addressing a broad

range of information requirements for both international and domestic programs

concerned with monitoring and inventory of renewable resources. Included were studies

of using remote sensing data for global crop production forecasts and estimates, for

rangeland and forest inventories, for land productivity analysis, for conservati~n practices

inventory, and for pollution detection and monitoring. The program which is ,currently

planned to continue through 1986 has produced many research results that are now in daily

operational use by USDA agencies.

3.2 SRS Role

Because of previous remote sensing research, SRS was selected by USDA as a major

participant in AgRISTARS, in both coordination and research roles. Under the research

role, SRS has been primarily concerned with applications research for domestic (US) crop

estimates. The major contribution of SRS to this program has been the continued

refinement of concepts developed during the 1972-77 period, and the proof of concept

- 16 -

demonstrated in the 1978 Iowa project. These refinements have been integrated into the

SRS methodology which has been used for quasi-operational Large Scale Application Tests

(LSAT) in 1980, 1981, 1982, and 1983. Under the AgRISTARS program, SRS plans were to

conduct LSAT's over two states (Iowa and Kansas) in 1980, with an additional two states

added each year until by the end of the program at least ten states would be included.

This goal was met for 1981 and 1982; however, due to budget restraints, only one new

state was added in 1983 and no new states are planned for 1984. In 1984, this quasi-

operational LSAT will be completed for the following seven states: Iowa, Kansas,

Oklahoma, Missouri, Colorado, Illinois, and Arkansas. Primarily corn and soybeans

estimates will be prepared for Iowa, Illinois, and Missouri; winter wheat estimates for

Kansas, Oklahoma, and Colorado; and cotton, soybeans, and rice estimates for Arkansas.

While results from these LSAT's have been quite varied due to many problems with the

Landsat program, with data delivery, with cloud cover, etc., the resulting estimates have

provided a useful input to the CRB in preparing more accurate crop area estimates for

these states. Operational procedures have been improved and costs reduced to the point..

where the methodology is cost-effective when compared to comparable results by simply

expanding the JES ground survey.

4. Future Use of Aerospace Remote Sensing

The future for aerospace remote sensing for domestic crop estimates is not clear at this

time. Primarily, because of uncertainties associated with the US land remote sensing

program. These include questions about whether the program (if there is one) will be

operated by the government or by private firms, about whether sensor and data delivery

improvements can be made, and whether or not total costs of a remote sensing system for

agricultural estimates will continue the downward trend established during the past

- 17 -

decade (SRS costs on a per state basis for applying this methodology has been reduced by

a factor of seven since 1975). Those involved with this research in SRS feel that the

potential of aerospace remote sensing for US crop estimates has been demonstrated and

plans are currently being made to refine the present methodology so that by the end of

AgRISTARS in 1986, SRS managment will be able to make a decision whether or not to

implement the technology on an ongoing operational basis. This decision will undoubtedly

be influenced not only by the US land remote sensing program (private or public) but also

by similar program.s in other nations, such as France with its "Systeme Probatoire d'

Observation de 1a Terre" (SPOT) satellite, and Japan with its Earth Resources Satellite

(ERS 1) (14)--satellites for land remote sensing which are planned for launch in 1985 and

1986, respectively.

Summary

Agriculture has long been identified as the field where applications of aerospace remote

sensing technology has the potential of making the largest impact in terms c;>fproducing

new information or improving existing information systems (16). The SRS of ~SDA over

the past decade has viewed aerospace remote sensing in the latter context viz., that it

represents a new source of data particularly on agricultural crops, which should be

exploited for improving the current agricultural statistics program. Consequently,

beginning soon after the launch of Landsat 1 in 1972, SRS has pursued an active research

program to develop methodology and data processing systems for using Landsat digital

MSS data for improving the existing US program of crop area estimates. After five years

of developmental studies, a major pilot test of the resulting methodology and data

processing system was conducted in 1978 for the entire State of Iowa. Based on results of

this test, which are described in this paper, SRS, under the AgRISTARS program,

- 18 -

developed a plan for a series of LSAT's beginning in 1980, designed to further demonstrate

the usefulness of Landsat data for crop area estimates. This plan included additional

research to refine and enhance the methodology for possible operational use after the end

of AgRISTARS in 1986. The 1983 LSAT covered seven (Kansas, Colorado, Iowa,

Oklahoma, Missouri, Illinois, and Arkansas) major crop states. Although results have been

variable, due to the quasi-research nature of the test, quality of Landsat data, and

problems with Landsat data delivery, the estimates have been significantly more precise

than those based on. ground surveys and have proven useful in establishing final end-of-

year estimates for crops grown in the seven states.

~lthough the LSAT's will be continued through 1986, the decision to proceed with an

operational remote sensing system for US crop area estimates will also depend on the

status of land remote sensing systems in operation, or anticipated, at that time. The

usefulness of aerospace remote sensing for crop area estimates has clearly been

demonstrated for Landsat MSS type data. It is expected that data from future sensors,

such as the one planned for the French SPOT satellite, will prove ,even more useful.,

However, costs of data from these future systems are still to be determined and their

ultimate use by SRS is still questionable. Current prospects would seem to indicate at

least a continued use by SRS of aerospace remote sensing data for information on US

crops. Expansion to other states and/or crops will be limited by scarce resources but will

probably continue to occur as research provides further gains in productivity for using

aerospace remote sensing.

- 19 -

Notes

IDirect expansion estimator (~Df.) and the regression estimator (YR)used by SRS

Y = Population total (for example, total area planted to corn)

YDf. = t.stimated total (using Jt.S ground data only)

where,

=Li-

i = I

n.N. ( L1

1j = 1

y .. >In.1 J 1

L = the number of land use strata in the state.

Ni = population number of sampling units (su) in stratum i.

ni = sample number of su in stratum i.

y.. = area planted to corn in the jth su of the ith stratum.IJ

Then the variance of YDEis:

j = I

where,

y. =1

L ~ N. -n.N.1 1 1= i- n.(n.-l) . N.

i = 1 1 1 1

n.i-1 /y .. n.

1 J 1

L - "i- (y .. -y.)i= 1 1 J 1

Yi;( = f.stimated total (using Landsat and Jf.S data)

L

= L N. Yi(reg)i = I 1

where,

Y.(reg) = y.+ b. (5<.-X.)1 1 1 1 1

- 20 -

and

bi = the estimated regression coefficient for the ith stratum when regressing ground

reported corn area on classified pixels for the ni suo

Xi = the population average number of pixels of corn per su in the ithstratum.

xi = the sample number of pixels of corn per su in the ith stratum.

Then, the estimated (large sample) variance of YR is:

L N~ N. -n. 2-l-r. n.V(YR) 1 1 1 __ 1

~1 - :t= l. (y .. -y.)i = 1

n. N. n. -2 j = 1 1 J 11 1 1

where,

ri= sample coefficient of determination between reported corn area and classified

. I f h . h .thcorn plxe s or t e ni su In tel stratum.n.

(£,1 (y .. -y.)(x .. -i.»~r~ = j=l IJ 1 IJ 1

1 n. n.(~1 (y ... _Y.)4:) ( £,1 (x .. _x.):t)i = 1 1 J 1 j = 1 1 J 1

Note that

So that lim V(YR·) = 0 as r.:t+ 1 for fixed n .•1 1 1

The relative efficiency of YR compared to YOE is defined as follows:

R. E. =

- 21 -

2R· .eglStratlon

As used in this paper, registration refers to the procedures for determining the geographic

location of each pixel in a given Landsat scene. It is the process of matching the Landsat

row-column coordinates with map latitude-longitude coordinates. The process involves

finding common points, such as road and/or rail intersections, on both Landsat images and

index maps. By digitizing map coordinates of these points, a mathematical

transformation can be calculated to predict the latitude-longitude coordinates of all

pixels in the Landsat scene. Further precision adjustments are made to accurately locate

the JES segments field-by-field in the Landsat scene. Over time this precision adjustment

has evolved from a largely manual approach to one that uses computer algorithms to

detect field edges within the Landsat data. The ultimate goal of registration is to locate

JES segments in the Landsat data to an accuracy of ! 0.5 pixel.

3Discriminant' Analysis - Classification

The basic use of discriminant analysis is to assign an observation to one of two or more

groups on the basis of its value. In this paper the observation is the "pixel" and its value is

the vector of measurements recorded by the MSS on Landsat. The multivariate

measurements represented by this vector were used to assign (classify) each pixel into a

crop or land use type (corn, soybeans, cities, water, etc.).

A sample of fields from the JES for each crop or land use type was selected and the MSS

data were summarized to obtain the distributions for each type. With four measurements

for each pixel the measurement space (MS) is four-dimensional. This largely precludes a

visual partitioning of the MS by simple visual means which may be appropriate for

separating observations into groups for one or two dimensions. With four-dimensional

- 22-

data, boundaries between groups become contour "surfaces" in the MS. These dividing

surfaces are constructed so that observations falling on the "surface" have equal

probabilities of being in either group. Those observations not on the dividing surface

always have a greater probability of being classified into the group for which the

observation is interior to the contour surface.

For SRS applications, discriminate functions have been computed by assuming the MSS

measurements are from a multivariate normal density distribution. Departures from

normality has not been a serious problem as long as precautions have been taken to insure

unimodal data. This has meant that in certain cases the data analyst has to recognize,

usually by using cluster analysis, that the same crop may fall into more than one group,

e.g., MSS data for late planted corn may follow an entirely different distribution than

corn planted early. After assuring that the normality assumption is approximately

satisfied, a multivariate normal distibution was estimated for each group. This was based

on a sample of fields from the JES for each crop or land use group. This sample was used

to calculate mean vectors and variance-covariance matrices for each crop group. Using

these sample statistics, the quadratic discriminate function was calculated. This function

was used to classify all unknown pixels into one of the crop groups for which the crop

mean vector is closest to the point based on the Mahalanobis distance. That is, the crop

for which the probability is highest.

References

1. "AgRISTARS Technical Program Plan," January 15, 1980, AP-JO-00628, JSC-I7153,

Lyndon B. Johnson Space Center, Houston Texas 77058.

2. Allen, R. D., "USDA Registration and Rectification Requirements." Proceedings of

the NASA Workshop on Registration and Rectification, Nevin A. Bryant, Editor, JPL

- 23-

Publication 82-83, Jet Propulsion Laboratory, California Institute of Technology,

Pasadena, California, June 1, 1982.

3. Amis, M. L., and others, "Evaluation of Large Area Crop Estimation Techniques Using

Landsat and Ground-Derived Data." AgRISTARS Report DC-LI-04051, JSC-17116,

LEMSCO-15763, March 1981.

4. Cardenas, M., M. Blanchard, M. Craig, "On the Development of Small Area

Estimators Using Landsat Data as Auxiliary Information," Research report, ESCS,

USDA, Washington, D. C., 1978.

5. Cochran, W. G., 1977. Sampling Techniques - Third Edition, John Wiley and Sons,

Inc., New York.

6. Cook, P. W., "Landsat Registration Methodology, USDA, SRS, 1972-1982." Research

report, Remote Sensing Branch, Statistical Research Division, SRS, May 1982,

Washington, D. C.

7. Craig, M., R. Sigman, and M. Cardenas, "Area Estimates by Landsat: Kansas 1976

Winter Wheat," Research report, ESCS, USDA, Washington, D. C., 1978•..

8. Gleason, C., and others, ''The Auxilliary Use of Landsat Data in Estimating Crop

Acreages: Results of the 1975 Illinois Crop Acreage Experiment." Research report

SRS-21, USDA, Washington, D. C., October 1977.

9. Hanuschak, G. A., R. D. Allen, and W. H. Wigton, "Integration of Landsat Data into

the Crop Estimation Program of USDA's Statistical Reporting Service, 1972-1982,"

Proceedings Eighth International Symposium, Machine Processing of Remotely Sensed

Data with Special Emphasis on Crop Inventory and Monitoring, Purdue University,

Laboratory for Applications of Remote Sensing, July 7-9, 1982.

10. Hanuschak, G., and K. Morrissey, "Pilot Study of the Potential Contribution of

Landsat in the Construction of Area Sampling Frames," Research report, SRS, USDA,

Washington, D. C. 1977.

- 24-

11. Hanuschak, G., and others, "Obtaining Timely Crop Area Estimates Using Ground-

Gathered and landsat Data." Technical Bulletin 1609, ESCS, USDA, Washington,

D. C., August 1979.

12. Huddleston, H. F., and W. H. Wigton, "Use of Remote Sensing in Sampling for

Agricultural Data," ISI!IASS Invited Paper, Warsaw, 1975.

13. Lindenlaub, J. C., "The lARSYS Software System--An Overview," 136 p. Purdue

University, laboratory for Applications of Remote Sensing, West Lafayette, Indiana.

14. Manual of Remote Sensing, Second Edition, Volume 1, Published by American Society

of Photogrammetry, 1983.

15. Ozga, M., W. Donovan, and C. Gleason, "An Interactive System for Agricultural

_ Acreage Estimates Using Landsat Data." Proceedings 1977 Symposium on Machine

Processing of Remotely Sensed Data, Purdue University, Laboratory for Applications

of Remote Sensing.

16. Remote Sensing, with Special Reference to Agriculture and Forestry, Committee on

Remote Sensing for Agricultural Purposes, Agricultural Board, National Research

Council, National Academy of Sciences, Washington, D. C., 1970.

17. Scope and Methods of the Statistical Reporting Service, U.S. Department of

Agriculture, Miscellaneous Publication No. 1308, Revised September 1983.

18. Sigman, R., and G. Walker, "Comparison of Auxiliary Variables for Satellite-Data

Regression Estimators in Crop Surveys."i

Contributed paper, 1983 American

Statistical Association Annual Meeting, Toronto, Ontario, Canada, August 1983.

19. Von Steen, D., and W. Wigton, "Crop Identification and Acreage Measurement

Utilizing Landsat Imagery," Research report, SRS, USDA, Washington, D. C., 1976.

- 25-